Re:Jailbreaking ChatGPT with Dan (Do Anything Now)
by digitailor on Monday February 06, 2023 @01:49PM
(#34682161)
That's more or less correct. My post was targeted to the investor segment of HN readership about labor automation, displacement, and re-valuation. The coder/techaesthete segment has focused exclusively on 1/7 sentences of my comment, as noted here, 20 min before your post, that quoted the 1/7 literally:
https://news.ycombinator.com/item?id=34681721
Which is pretty cool, actually
Re:Jailbreaking ChatGPT with Dan (Do Anything Now)
by Dylan16807 on Monday February 06, 2023 @03:44PM
(#34684064)
If you make a very strong point in your first sentence, of course people will focus on that. And while the rest of your post can stand alone, you made that sentence part of the foundation of your argument. "going to show" "in other words" "so" So arguing against that point is relevant to notably more than 1/7 of your post.
Re:Jailbreaking ChatGPT with Dan (Do Anything Now)
by digitailor on Monday February 06, 2023 @04:40PM
(#34684812)
Yup, you’re dead on, that’s how "engagement" tends to work, but hook is not also line and sinker, no? So advanced speech generation models are now having to account for engagement— contextually— as well. It’s all getting much more refined, somewhat rapidly, but not necessarily truly usefully
Eh. Other than the penalty being fake, the original DAN didn't have anything like that at all and got similar results. It was just a pep talk and a command to give the normal answer and then the DAN answer.